Language models, which are components of systems such as automated speech recognition (ASR) systems, assign probabilities to sequences of text. ASR systems transcribe utterances into a series of computer readable sounds, which are then compared to a dictionary of words in a given language. The relative probabilities associated with several candidate sequences of words can be used to help select the most likely transcription of an utterance.